pulsatrix::lrp_composite Namespace Reference

Zennit 1.0.0's composite presets (zennit.composites), mapped onto pulsatrix modules by type: LinearModule = torch Linear, Conv2DModule = torch Conv2d; every other module gets the epsilon rule (the pass-through modules – ReLU, Flatten, Dropout, MaxPool – ignore it, as Zennit's Pass rule / plain gradient does). More...

Namespaces
namespace	detail

Functions
LRPComposite	epsilon_plus (float epsilon=1e-6f)
	Zennit EpsilonPlus: Epsilon for Linear, ZPlus (AlphaBeta 1, 0) for Conv2D.
LRPComposite	epsilon_alpha2_beta1 (float epsilon=1e-6f)
	Zennit EpsilonAlpha2Beta1: Epsilon for Linear, AlphaBeta(2, 1) for Conv2D.
LRPComposite	epsilon_gamma_box (float low, float high, float gamma=0.25f, float epsilon=1e-6f)
	Zennit EpsilonGammaBox: ZBox(low, high) for the first Conv2D layer (lowest index), Gamma(gamma) for every other Conv2D, Epsilon for every Linear.

Detailed Description

Zennit 1.0.0's composite presets (zennit.composites), mapped onto pulsatrix modules by type: LinearModule = torch Linear, Conv2DModule = torch Conv2d; every other module gets the epsilon rule (the pass-through modules – ReLU, Flatten, Dropout, MaxPool – ignore it, as Zennit's Pass rule / plain gradient does).

Note

Every Epsilon entry sets epsilon_bias_in_denominator = true: Zennit's Epsilon rule puts the bias in z, so matching Zennit numerically requires it (this codebase's own default epsilon rule leaves the bias out to be conservative).

Function Documentation

epsilon_alpha2_beta1()

LRPComposite pulsatrix::lrp_composite::epsilon_alpha2_beta1 ( float  epsilon = 1e-6f )

inline

Zennit EpsilonAlpha2Beta1: Epsilon for Linear, AlphaBeta(2, 1) for Conv2D.

epsilon_gamma_box()

LRPComposite pulsatrix::lrp_composite::epsilon_gamma_box ( float  low, float  high, float  gamma = 0.25f, float  epsilon = 1e-6f  )

inline

Zennit EpsilonGammaBox: ZBox(low, high) for the first Conv2D layer (lowest index), Gamma(gamma) for every other Conv2D, Epsilon for every Linear.

Note

As in Zennit 1.0.0, whose first_map holds only Convolution: a Linear is never ZBox'd, even when it is the first layer, so on a Conv2D-free network this preset is Epsilon on every layer (checked against Zennit in tests/lrp_reference_test.cpp).

Remembers the first Conv2D it has seen since the last index-0 call (see LRPComposite's calling order).

epsilon_plus()

LRPComposite pulsatrix::lrp_composite::epsilon_plus ( float  epsilon = 1e-6f )

inline

Zennit EpsilonPlus: Epsilon for Linear, ZPlus (AlphaBeta 1, 0) for Conv2D.